Multi-modal Market Manipulation Detection in High-Frequency Trading Using Graph Neural Networks

Abstract

This paper proposes a novel multi-modal graph neural network framework for detecting market manipulation in high-frequency trading environments. The framework integrates diverse data sources through sophisticated fusion mechanisms and employs attention-based graph neural networks to capture complex trading patterns. Our approach constructs dynamic trading networks that encode temporal and structural dependencies, enabling the detection of subtle manipulation strategies. The model architecture incorporates multiple attention layers for feature selection and cross-modal information fusion, achieving superior detection performance compared to traditional methods. Experimental results on real-world high-frequency trading data from major exchanges demonstrate the framework's effectiveness, reaching 98.7% accuracy in manipulation detection while maintaining low latency (8.3ms average processing time). The model exhibits robust performance across various market conditions and manipulation patterns, with precision and recall rates exceeding 97%. Through comprehensive case studies and interpretability analysis, we validate the framework's ability to identify and explain complex manipulation strategies while providing insights for regulatory compliance. The proposed approach advances state-of-the-art market surveillance technology, offering a scalable solution for real-time manipulation detection in modern financial markets.